Floods, which are becoming more frequent due to climate change, are potentially threatening to the high population density areas and complex infrastructures in urbanized zones and cause huge social, economic and environmental damages. The most significant challenge with the traditional methods (physics-based) of flood prediction is their speed of simulation, which makes it difficult to provide timely predictions for the occurrence of urban flood events. Recently, by analyzing big datasets such as weather patterns, terrain characteristics and historical flood records, machine learning (ML) models have shown promising results in improving the accuracy of flood predictions and flood maps. This work presents metadata analysis aiming to understand recent efforts using ML and specifically, deep learning (DL) approaches (e.g., Decision Trees, Support Vector Machines, Convolutional Neural Networks, Long-Short Term Memory, etc.) to predict timing, extents, and urban damages caused by flash floods. The literature highlights a wide range of input data, but the most common are rainfall, slope, elevation and distance from river and roads. Model performance metrics, that have been used the most are precision (0.7-0.98), accuracy (0.64-0.98) and AUC (0.69-0.99). Additionally, out of 112 research papers, 40 are from China, reflecting the country’s significant focus to improving flood prediction models.